Given this was about a decade ago, there are probably better predictive models than whatever was in the demo he showed me, but I wouldn’t be surprised if their model was a Markov chain where the size of each next-letter option was a function of probability.
(Memory haze: I was more excited by the hyperbolic space than the word model when he showed it to me, and it was c. 2008)
Even if I change protecting to projecting I don't get it.
Well imagine if you would like to draw as many intervals of any given shape but with the same length on a line.
Let's start out with a line. Then you can only have so many symbols given by the height.
To continue if we instead where to space them out over the circumference of a circle, we would on the same width have more symbols.
If we go even further and give the circle jagged edges that go in and out. Theoretically we have an infinite length on the line, but resolution on screen and eyes sets an stop to that.
So basically making use of the area of the screen instead of only a line.
Did I make more sense to you? The fractal thingy is just an experiment and not the main point of the comment so I gave a very brief description of it. It may very well be that my capabilities of describing it in proper mathematical is a bit lacking.
Will have a demo of it within a few days or so if you are interested keep your eyes open.
I don't know if you are familiar with the Dasher project for text input, but as of know I'm trying to improve on that work partly by improving on how many letters are available simultaneously by protecting the line of text upon a fractal surface. Something that should be a more efficient use of a 2d surface, theoretically infinitely so.
As far as autocomplete is concerned my approach is to try to do exactly this but on a character basis. I think this can lead to some interesting advantages, for example different dialects gives rise to words that not always conform to dictionary specifications.
The next level would be to go one step higher so to speak. If we imagine Markov Chain on letters as the first level and said chains level, I'd say that the third level in our hierarchy would be to apply markov chain on groups of words grouped by proximity in a word2vec space.
Having markov chains working on groups of word2vec words would give us a statistical analogy of grammar. However without having to implement it programatically, something that inevitably would lead to missed corner cases and if not that a too strict algorithm that would hinder intentional abuse of grammar by purpose.
Maybe this is already being implemented, as it to me seems as the logical next step. Anybody got any info on this?